South Africa’s AI ambitions face a looming power constraint

Warren Olivier | image supplied

As South Africa prepares for further data centre expansion, planned coal-plant retirements could put renewed pressure on the grid, raising questions about the cost and reliability of running AI workloads locally.

Data centres face growing electricity demand

New data centres planned for Cape Town are expected to require a combined 580MW, equivalent to more than a third of the city's peak electricity demand of 1 676MW. At the same time, Eskom's planning points to roughly 5GW of coal-fired generation capacity retiring from 2029.

This convergence of rising demand and retiring generation capacity could create a new constraint for businesses looking to expand their use of AI.

Some organisations are bringing computing workloads closer to home to improve cost predictability and control over their data. However, access to the hardware needed to run AI models is only part of the challenge. Graphics processing units (GPUs), which handle much of the computation involved in AI, cannot operate without a dependable electricity supply.

The question for businesses is therefore not only how much computing capacity they can secure, but where they can afford to run it and how they will maintain operations if electricity becomes more constrained.

AI introduces new cost uncertainties

The cost of running AI workloads can be difficult to predict. Cloud providers use different pricing models, while increased use of more capable AI models can drive up computing demand and expenditure.

For finance teams accustomed to forecasting conventional cloud costs, this creates an additional budgeting challenge. As AI adoption grows, businesses may find that their actual computing costs diverge significantly from their original estimates.

Electricity constraints could compound the problem. If power becomes more expensive or less reliably available, businesses running AI workloads in local data centres could face higher operating costs at the same time as they need computing capacity to remain available.

This makes cost management and operational continuity closely linked. An organisation that brings its AI workloads closer to home to gain greater control over spending could still face unexpected costs if the electricity supply becomes constrained.

The limits of diesel backup

Backup generation can help data centres maintain operations during power interruptions, but diesel generation has practical and financial limits.

The backup capacity available at individual data centre operators is not always publicly disclosed, making it difficult to assess how long facilities could sustain operations during a prolonged disruption. Supplying hundreds of megawatts of demand across multiple large facilities would also present substantial logistical and fuel-supply challenges.

For context, Eskom's open-cycle gas turbine fleet, which uses diesel to provide additional electricity to the national grid, has a combined capacity of less than 3GW. Although these turbines serve a different purpose from data centre backup systems, the comparison illustrates the scale of generation capacity required to support electricity-intensive operations.

Businesses may also need to account for additional conditions in their service agreements as power constraints evolve. These could affect continuity arrangements, recovery planning and the assumptions underpinning decisions to move workloads from international cloud environments to local infrastructure.

Diversifying where AI workloads run

Local computing capacity can offer advantages, including greater control over data location and reduced dependence on overseas infrastructure. However, concentrating workloads in one location can introduce another form of risk if that location faces electricity shortages or rising energy costs.

A more resilient approach is to avoid depending entirely on a single computing environment. Businesses can assess whether different workloads should run in local data centres, international cloud environments or other available facilities, depending on their operational requirements, data governance obligations and costs.

This approach requires planning for more than a simple transfer of workloads. Organisations need to consider whether their applications can operate across different environments, how quickly they can move processing capacity, and whether their data can remain accessible without expensive or time-consuming transfers.

Flexibility becomes a business continuity requirement

One possible approach is to keep data in its existing location while directing computing tasks to suitable infrastructure elsewhere. Where systems and data governance requirements allow it, businesses could use alternative providers to accommodate spikes